Calibrating vascular abundance for detecting future climate changes in Oregon and Washington, USA

a * b b Timothy J. Brady : , Vicente J. Monleon , Andrew N. Cray

ABSTRACT

We propose using future abundances as indicators of future climate in a way analogous to the reconstruction of past environments by many palaeoecologists. To begin monitoring future short- term climate changes in the forests of Oregon and Washington. USA. we developed a set of transfer functions for a present-day calibration set consisting of climate parameters estimated. and abundances measured. at 107 USDA Forest Service FIA(Forest Inventory and Analysis) Phase 3 plots. For each plot. we derived climate estimates from the Daymet model database. and we computed species abundance as quadrat frequency and subplot frequency. We submitted three climate variables (mean January temperature. MJAT; mean July temperature. MJUT; and mean annual precipitation transformed to natural logarithms. MANPt) to canonical correspondence analysis (CCA). and verified their importances in structuring the species frequency data. Weighted averaging-partial least squares regression (WA-PLS) provided the means for calculating six transfer functions. In all cases. based on performance statistics generated by leave-one-out cross-validation. we identified two-component WA- PLS models as the most desirable. The predictive abilities of our transfer functions are comparable to. or better than. those reported in the literature. probably due both data quality and statistical considerations. However. model overfitting as a result of spatial autocorrelation remains a possibility. The large errors associated with our MJAT transfer functions connote that even the highest amount of change in mean January temperature predicted for Oregon and Washington for 2010-2039 would be indistinguishable from current conditions. The higher predictions indicate that our MJAT transfer functions may be able to track climate changes by the 2040s. Our MJUT transfer functions can detect change in mean July temperature under the highest projection for 2010-2039. Our MANPt transfer functions will be of limited use until the 2070s. given the predictions of only slight changes in mean annual precipitation during the early part of the twenty-first century. Our MJAT and MANPt transfer functions may prove useful at the present time to verify relative climatic stability. Because the predicted climate values sometimes deviate substantially from the observed values for individual plots. our transfer functions are appropriate for monitoring climatic trends over the entire Pacific Northwest or large regions within it. not for assessing climate change at individual plots. © 2009 Elsevier Ltd. All rights reserved.

1. Introduction species as indicators potentially offers very high-resolution data on local climate where instrumentation is unavailable (e.g., in remote Climate Significantly affects the growth rates and equilibrium mountainous locations). Given the recent and expected future rapid abundances of populations of vascular . and therefore. it climate change (e.g ., IPCC, 2007). decision makers demand high influences their evolutionary histories through natural selection quality quantitative descriptions of local climate at successive time (Enright. 1976; Woodward, 1987, 1992; Woodward and Williams, intervals (Hartmann et al., 2002; Miles et al., 2006). Vascular plant 1987). In effect, the genomes of vascular plant species contain indicators may help meet this need, but they may also provide information about past and present climates. If properly calibrated, information regarding the biological impact of changing climate. an assemblage of vascular plants should provide a quantitative Outside of palaeoecology, few researchers have used vascular signal of site-specific climatic conditions. The use of vascular plant plants as indicators for obtaining quantitative estimates of local climate variables: Karlsen and Elvebakk (2003) mapped local temperature variations in East Greenland using an "Index of Therrnophily'', a combination of a plant species' summer temper- ature needs (determined by geographical distribution), its local abundance, local habitat diversity, and the local proportion of plots, which represent a systematic subset of the inventory plot unproductive land. More recently, Garbolino et al. (2007) grid that spans the forestlands of the USA. The density of Phase 3 determined the climate affinities of nearly 2000 vascular plant plots is about one per 390 km2. Each plot consists of a cluster of species in France based on correlations between the geographical four 168 m2 non-overlapping subplots, arranged within a 1 ha distributions of reIeves and meteorological stations. For each of six circle. Each subplot contains three 1 m2 quadrats. Field crews climate variables, they calculated an affinity as a function of a recorded the presence and cover of each species with heights of up species' maximum frequency and the dispersion of its frequencies to 1.8 m on each quadrat. They subsequently searched the subplots around the maximum. Using vascular plants as indicators, the for up to 45 min to find additional vascular plant species of any authors produced detailed climate maps of France. height that occur outside the quadrats. Vegetation sampling of Since the 1970s, many palaeoecologists have employed Phase 3 plots occurs during June, July, or August, and, if fully vascular plants as indicators to derive quantitative climatic implemented, remeasurement would take place on the standard information about local or regional late-Quaternary environments inventory cycle (10 years in the western USA). Stapanian et al. (recent examples include Finsinger et aI., 2007; Li et al., 2007; (1997) and Schulz et al. (2009) provide detailed information about Neumann et al., 2007; Wilmshurst et al., 2007; St.jacques et al., the design and sampling of FIA Phase 3 plots. 2008). These researchers have sought to develop transfer func- Cover values reflect the vagaries of phenology, weather tions, multivariate equations that permit the estimation of ancient conditions, field crew biases, and actual abundance (Gray and climate parameters from data on fossils (Birks, 1995, and Azuma, 2005; Smith et al., 1986; Schulz et al., 2009). As an references therein). Calibration sets, which include spatially alternative, we computed two frequency measures for each species correlated present-day climatic conditions and pollen grain/ present on a given plot. For species i, Fqi is quadrat frequency and pteridophyte spore/phytolith counts, specify the transfer func- FSi is subplot frequency: tions. Formally, C (a matrix of n standardized observations, one at each of n locations on q climate variables), X (the corresponding matrix of standardized observations on p biological variables at each ofthe same n locations), B (a matrix of transfer functions, one for each of the q climate variables, with each comprised of p elements), and 0 (a matrix of random noise, with n parts for each of the q climate variables) are related through a linear model: Both Fqi and FSi vary from 0 to 1. Quadrat frequency and subplot frequency provide rough yardsticks of abundance, not just distribution, of a species within an FIAPhase 3 plot. Adult survival, B acts as a set of transfer functions, because observations on the p fecundity, within-plot dispersal, and the survival of recruits dictate biological variables render estimates of the q climate variables. the frequency of a vascular plant species. While they are not The use of vascular plants to monitor climate carries the synonymous with population density, Fqi and FSi convey informa- assumption that their populations are always in dynamic equilibri- tion about the proportion of a plot occupied by a species. A plant um with climate (Davis and Botkin, 1985; Webb, 1986; Prentice species with a large population, as measured by cover value, et al., 1993): the list of species in a community and their properties should occupy a larger proportion of a plot than a species (e.g., abundances) match those allowable under the prevailing represented by a small population (Andrewartha and Birch, 1954; climatic regime. If the biological response lags significantly behind Krebs, 1985). the pace of climate change, then this assumption is unjustified. Davis Like all frequency measures, Fqi and FSi depend on plot size. and Botkin (1985) showed, through simulation studies, that short- Thus, we excluded from consideration any plot comprised of less term (i.e., decadal) climate changes may have little immediate effect than 100% accessible forest, as it would have sampled fewer than on the adult survival of long-lived tree species. Even annuals and 12 quadrats and fewer than four subplots. short-lived perennials may be little affected by climate change as For construction of the calibration set, we obtained data from long as the buffering capacity of the canopy persists (Smith, 1965; 107 fully forested FIA Phase 3 plots sampled in 2001, 2004, and Davis and Botkin, 1985). Regular disturbance may be necessary to 2005 distributed across Oregon and Washington (Fig. 1). We guarantee that short-term climate change will bring about a calculated quadrat frequencies for 538, and subplot frequencies for measurable species response. 698, vascular plant species on these plots. In view of recent Palaeoecologists use the past remains of vascular plants as findings (Lischke et al., 2002; Miller et al., 2008; Payette, 2007; indicators of past climate. Transfer functions provide the means of Svenning et al., 2008; Valsecchi et al., 2008), we expect that any lag translating past biology into quantitative estimates of past climate. in response (a delay in adjustment of species frequency on a plot) We propose that foresters and environmental scientists adopt a to climate change will depend upon the complex interplay of many parallel approach: use future data on vascular plants as indicators of factors such as species autecology (including vagility), forest stand future climate. Transfer functions will furnish the link between structure (resulting from intra- and interspecific interactions as future biology and future climate. In the present study, we took the well as disturbance regime), soil development, and geographical first step in establishing a program for monitoring future short-term impediments to migration (e.g., rugged terrain and habitat climate changes in the forests of Oregon and Washington, USA, fragmentation). The immediacy of response may vary among based on vascular plant abundances: We developed a set of transfer species on the same plot, and even among plots for the same functions for a calibration set consisting of a few estimated climate species. We cannot possibly predict the speeds of species parameters and species abundances measured at 107 locations. responses, especially since most species analyzed here are poorly known ecologically. Therefore, we included a broad array of plant 2. Materials and methods functional types in this study (e.g., annuals, short-lived perennials, ruderals, and long-lived perennials) to increase the likelihood of 2.1. Data on vascular plant abundance obtaining rapid and significant changes in frequency by at least some species on each plot following short-term climate change. We derived vascular plant abundances from inventory data for Unlike conventional palaeoecological practice, we did not exclude USDA Forest Service FIA (Forest Inventory and Analysis) Phase 3 rare species from the calibration set. While this maneuver engenders the risk of producing poorly performing transfer somewhere in the study area. Of course, this assumes that climate functions, it may increase the sensitivity of our analysis. change will not reduce the fully forested condition of a plot. We sought to use vascular plant species as indicators of four 2.2. Climate data general climate variables: (1) mean annual temperature oC; here symbolized MANT), (2) mean january temperature oC; here To obtain valid approximations of local climatic conditions in symbolized MjAT), (3) mean july temperature oC; here symbolized Oregon and Washington after 2005, the calibration set must sample MJUT), and (4) mean annual precipitation (mm; here symbolized the full range of future climates in these two states (Sachs et al., MANP). We obtained a plot-specific 18-year (1980-1997) average 1977; Howe and Webb, 1983). Given the pronounced climatic for each of these variables by intersecting plot locations with data diversity of Oregon and Washington (climate varies from maritime from Daymet (Thornton et al., 1997; database at http://www.day- to continental, with strong orographic and rainshadow effects met.org/), a high-resolution (1 km grid) model developed for the caused by numerous tall [>3000 m] mountain ranges; Franklin and conterminous USA from elevation data, topographic indices, daily Dyrness, 1988), we expect that most future shifts in climate will observations from ground-based meteorological stations, and result in conditions at one location that match those found today statistical algorithms. The temperature variables (MANT, MJAT, and MjUT) were marked ecological diversity of the Pacific Northwest (Franklin and roughly normally distributed to slightly negatively skewed. Dyrness, 1988), represent the longest compositional gradients However, MANP was severely positively skewed, due to very high reported in the literature for vascular plant transfer function precipitation values near the Washington coast. We transformed studies, They strongly suggest that species responses to MJAT, the mean annual precipitation data to natural logarithms (MANPt MJUT, and MANPt are non-linear. symbolizes the transformed variable) to produce an approximately We used the fitting procedure of Huisman et al. (1993) to decide normal distribution. While the statistical techniques used in this whether the relationships between species abundance and climate study do not require a normal frequency distribution over plots for are best approximated by a symmetrical unimodal model or by one a climate variable, a reduction in skewness improves transfer of three other non-linear functions, The alternatives (so-called HOF function performance. Table 1 gives descriptive statistics for the models) are elements of an hierarchical set: a skewed unimodal frequency distributions of MANT, MANP, MJAT, MJUT, and MANPt. response (model V), a symmetrical unimodal response (model IV), a monotonic increase or decrease with plateau (model III), a 2.3. Climate variable selection and evaluation monotonic increase or decrease (model II), and a flat response (model I), The latter connotes that species abundance is unrelated In the spirit of parsimony and to comply with the requirements to the environmental variable, Simplicity increases from model V of the statistical techniques used here, we sought to minimize to model I, because the number of estimated parameters declines redundancy (multicollinearity) among the climate variables (although models III and IV contain the same number of included in our analyses. We calculated variance inflation factors parameters). Oksanen and Minchin (2002) produced an algorithm (VIFs: Marquardt, 1970) for all climate variables with JMP 4.0.2 for model selection that proceeds by fitting the most complex (SAS Institute Inc., 2000). The VIF for MANT was >10, which model (V) to the species data using maximum likelihood (with a according to Belsley et al. (1980), indicates a high degree of Poisson error structure), and then by reducing model complexity multicollinearity. Therefore, we eliminated MANT from further (to IV, then to III, then to II, and finally to I) until deviance increases analysis. This reduced the VIFs for the remaining variables to <4. significantly (an F-ratio test fails, where P?: 0.05). According to The methods employed here assume that the frequencies of their routine, examination of model III occurs only after rejecting vascular plants are significantly correlated with climate variables, model IV over Model V. Therefore, model III may, in fact, represent but only weakly linked to non-climatic factors (Birks, 1998, and a superior depiction of the response of some species allocated to references therein). We used canonical correspondence analysis model IV. We implemented this procedure for each species (CCA: Ter Braak, 1986, 1987), to evaluate the relationship between detected on at least ten plots and the three selected climate the biological and climate variables in the calibration set, and to variables with HOF 2.3, an MS-DOS program (Oksanen and determine which variables we should include in the transfer Minchin, 2002; available at http://cc.oulu.fi/~jarioksa/) for the functions. CCA performs best when species abundance (here, quadrat and subplot frequency datasets. For both datasets, non- quadrat frequency or subplot frequency) is a symmetrical linear relationships to climate were statistically significant for unimodal function of each environmental variable (here, each of most species, and the symmetrical unimodal function was the the climate parameters). most common response type (Table 2). Species responses typically appear linear when gradients are We carried out CCA using the species data and three climate short, but they tend to assume non-linear forms when gradients variables with PC-ORD 5.0 (McCune and Mefford, 2005). As Palmer are long (Leps and Smilauer, 2003; 0kland, 1986; Ter Braak and (1993) found that CCA will not produce an artificial arch effect. a Prentice, 1988; Ter Braak, 1995). We used detrended correspon- detrended version of CCA was unnecessary. We completed several dence analysis (DCA; Hill and Gauch, 1980), to estimate the CCA runs for each species dataset (quadrat frequencies and subplot gradient lengths within the calibration set with PC-ORD 5.0 frequencies). During a run. the program rescaled each plot score to (McCune and Mefford, 2005). The program downweighted rare a mean of zero and a variance of one, it calculated species scores as species, divided each axis into 30 segments, and evaluated the weighted means of plot scores. and it computed ordination axis ability of DCA to recover compositional gradients by calculating, scores as linear combinations of the constraining climate variables. 2 for each axis, a coefficient of determination (r ) between For a given species dataset, we ran CCA for each climate variable Euclidean distance in the ordination space and relative Euclidean individually to assess its statistical significance (the program distance in the original species frequency space (McCune and executed a Monte Carlo test with 998 randomizations of the null Grace (2002) recommend the use of relative Euclidean as a hypothesis that no linear relationship exists between the matrix of distance measure for the original space under DCA). The species frequencies and a column matrix of climate values). We calculated gradient lengths for the quadrat and subplot frequen- then performed an ordination with all significant climate variables cies were over seven standard deviation units for the first two DCA to learn how they structured the species frequencies. To help axes, and the two axes explained 0.40 and 0.55 of the variance for evaluate the effectiveness of the latter, the program provided, for 2 quadrats and subplots, respectively. These figures are large for each ordination axis, a coefficient of determination (r ) between species-rich data with numerous zero abundance values. The Euclidean distance in the ordination space and relative Euclidean large gradient lengths, which are not unexpected given the distance in the original species frequency space (McCune and and other environmental characteristics (Franklin and Dyrness, 1988). Therefore, many variables, some behaving more-or-Iess randomly, others acting systematically, undoubtedly determine the species frequencies within a given plot. CCA will identify which of the climate variables structure the frequency data. During a WA- PLS run, all of these variables other than the one explicitly considered (and surely unidentified environmental variables as well) provide structured noise of potential use in improving model fit. For each climate variable deemed important by CCA,we used C2 1.5.0 (Juggins, 2007) to perform two separate WA-PLS runs, one for each set of species data (quadrat frequencies and subplot frequencies). We generated performance statistics for each WA- PLS run through leave-one-out cross-validation. Spatial autocorrelation within the calibration set (e.g., similar climate or species frequencies at nearby plots), if it affects the residuals (contained within 0 of Eq. (1 )), may lead to model overfitting (Telford and Birks, 2005, 2009; Telford, 2006). To check for spatial autocorrelation, we carried out a Mantel test (Mantel, 1967) for each of three datasets (quadrat frequencies, subplot frequencies, and plot-specific values for the selected climate variables) through PC-ORO 5.0 (McCune and Mefford, 2005). During a run, the program first converted the dataset to a Euclidean distance matrix. The program then performed a Monte Carlo test to evaluate its relationship to a matrix of inter-plot horizontal distances in meters (we ignored elevational variations between Grace (2002) recommend the use of relative Euclidean as a plots) by permuting (1000 times) the rows and columns of one of distance measure for the original space under CCA). the matrices. Rejection of the null hypothesis of no relationship between the two matrices provides evidence that the dataset 2.4. Calculation of transfer functions contains spatially autocorrelated information.

We used weighted averaging-partial least squares regression 3. Results (WA-PLS; Birks, 1998, and references therein) to estimate the transfer functions (B in Eq. (1)). A single WA-PLS run yields a 3. 1. Importance of climate variables transfer function for only one climate variable by constructing a model comprised of one or more components. WA-PLS, like CCA, The results of CCA demonstrate that MJAT, MJUT, and MANPt assumes that species-climate response curves are unimodal and substantially influence the frequencies of vascular plants (both symmetrical. It produces an initial component as a set of quadrat frequencies and subplot frequencies) in the calibration coefficients, or weighted averages of the species optima (means, dataset. The outcomes for quadrat and subplot frequencies were or distribution centers) with respect to a given climate variable. nearly identical. For brevity, we present the results only for the WA-PLS estimates the optima by inverse linear regression, i.e., it former. selects the optima so that their weighted averages best predict the We found no evidence of multivariate outliers duri ng any of the values of the climate variable in the calibration set. Higher runs, so we used the full dataset (538 species and 107 plots). components use the residual structure in the data to improve the CCA performed separately for the three relatively non-multi- estimates of the optima. Identification of the model (number of collinear climate variables indicated that MJAT, MJUT, and MANPt components) that gives the best transfer function requires an explain 2.9%, 2.4%, and 3.2% of the variance among species in the examination of performance statistics generated by leave-one-out space defined by a single climate variable (ratio of the eigenvalue cross-validation (jackknifing). In general, a low RMSEP (root mean for the first and only canonical axis to the total explainable square error of prediction), a low maximum bias, no appreciable variance), respectively. We are not surprised by such small decline in r2 (coefficient of determination), and a small number of percentages given the large number of species, numerous zero components characterize the preferred model. WA-PLS has proved abundance values, and noise contained in the quadrat frequencies to be a robust method for building transfer functions. WA-PLS is database. Nevertheless, all three climate variables exert statisti- well suited for analyzing datasets comprised of presence/absence cally significant influences on species frequency: For each CCArun, or compositional species records with large numbers of species a Monte Carlo test showed that the proportion of randomized runs and many zero abundance values, considerable (but structured) with an eigenvalue greater than or equal to the eigenvalue actually noise, long compositional gradients (high beta diversity), and associated with the first axis (P) is 0.001. symmetrical unimodal response curves. Table 3 summarizes the results of CCA as constrained by all The characteristics of our data help justify the use of WA-PLS for three (statistically significant) climate variables for quadrat transfer function construction: as argued previously, the frequency frequencies. The first of three canonical axes captured most of data used here reflect species abundances within plots. Neverthe- the variation in the ordination, with a coefficient of determination 2 less, they represent mathematical summaries of presence/absence (r ) of 0.22. Axis 1 is a precipitation gradient, with an almost data. The quadrat frequency database contains 538 species with perfect positive intraset correlation with MANPt. MJAT displays a 96% of the abundance values being zeros. The subplot frequency moderately high (and positive) correlation with the same axis, database includes 698 species with 95% zero abundance values. possibly revealing a gradient of "continentality". However, The plots included in the calibration set sample a broad array of inspection of ordination scores revealed that species near one geographical locations in a region known for its diversity in climate end of axis 1 tolerate scarce precipitation, but not necessarily cold winters (continental climates have low MANPt and low MJAT). Species near the other extreme of axis 1 enjoy copious precipita- tion, but not necessarily warm winters (maritime climates have high MANPt and high MJAT). 2 Axis 2 is a relatively weak climate gradient, with an r of 0.06. graminoid appearing on the right side of the ordination and close Both MJAT and MJUT are highly (and positively) linked to axis 2, to axis 1, occurred on a single plot in the wet, maritime-influenced indicating that it is a temperature gradient. western hemlock (Tsuga heterophylla) forest of the Coast Range in The small and negative r2 of axis 3 (-0.07) may reflect the Washington, Antennaria rosea, which appears on the left side of the optimization of some criterion unrelated to the percentage of ordination and close to axis 1, inhabits the dry Ponderosa pine variance extracted (McCune and Grace, 2002). The magnitudes of (Pinus ponderosa) and Douglas-fir (Pseudotsuga menziesii) forests of all intraset correlations with axis 3 are small to modest, indicating northeastern and southeastern Washington. At the bottom of the that the third axis is of no use in assessing the roles of the selected figure and close to axis 2, is a set of seven species (Antennaria climate variables in structuring the quadrat frequencies data. lanata, aureus, Erigeron lonchophyllus, Festuca viridula, The CCA ordination biplot graph (Fig. 2) provides visual details Potentilla drummondii, Ranuncultls verecundus, and Veronica ser- about the climatic correlates of the two useful ordination axes. pyllifolia) found on a plot in the high-elevation lodgepole pine Each climate variable arrow is based on the interset correlations (Pinus contorta) forest in the North Cascades of Washington where and eigenvalues; the length and angle indicates the strength of the temperatures are relatively low. Rupertia physodes, which appears correlation between the variable and axis. Luzula arcuata, a near the top of the figure and close to axis 2, occurs on only one plot in the warm Douglas-fir forest of the Klamath Mountains in southwestern Oregon.

3.2. Transfer functions

The abilities of WA-PLS to predict the three climate variables shown to be important by CCA (MJAT, MJUT, and MANPt), as assessed by leave-one-out cross-validation, were quite similar for both species databases (quadrat frequencies and subplot frequen- cies). Therefore, we present the complete results only for the WA- PLS runs based on quadrat frequencies. We removed no species or plots as supposed outliers from the quadrat frequencies dataset (538 species and 107 plots). Table 4 gives performance statistics. For all three WA-PLS runs (one for each of three climate variables), we identified two-component models as the most desirable. In all three cases, the two- component model yielded the minimal RMSEP (root mean square error of prediction), the minimal maximum bias, and the maximal r 2 (coefficient of determination). Fig. 3 permits a graphical evaluation of the performance of the two-component WA-PLS model for each climate variable. The scatter plots clearly reveal pronounced systematic bias in the predictions for all three of the climate variables (the models overestimate the climate values near the lower ends of the ranges and underestimate them near the upper ends of the ranges) a common shortcoming of WA-PLS (Ter Braak and Juggins, 1993; Birks, 1998). Only minor differences in performance, as indicated in Table 5, distinguished the results of WA-PLS (two-component models) for the quadrat frequencies and subplot frequencies datasets. For The Mantel tests revealed that, indeed, the calibration set shows MJAT, RMSEP (root mean square error of prediction) was about 12% significant spatial autocorrelation. The Monte Carlo tests yielded P lower for subplot frequencies. The maximum biases associated values of 0.012, 0.001, and 0.001, respectively, for quadrat with subplot frequencies were approximately 17% lower for MJUT, frequencies, subplot frequencies, and climate data (MJAT, MJUT, but around 10% higher for MANPt. and MANPt). For each dataset, we were compelled to reject the null specific species abundances, i.e., they are unaffected by pollen grain/spore/phytolith production rates, transport to sedimentary basins, and taphonomic processes. The scatter plots (Fig. 3) show that the transfer functions successfully predict the overall trends inherent in the observed climate values within the calibration set. However, the relation- ship between observed and predicted values is not perfect: many residuals deviate substantially from zero. Such deviance is due to systematic bias (expressed by maximum bias) and random variability (embodied in RMSEP).The former includes the tendency of WA-PLS to produce overestimates near the lower ends of the climate ranges and underestimates near the opposite ends. The latter results from plot-specific non-climatic factors (natural disturbance and secondary succession, soil characteristics, anthro- pogenic forces, and historical peculiarities) that influence species frequencies. The discovery of statistically significant spatial autocorrelation within the calibration set constitutes a warning that the transfer hypothesis of no correlation with inter-plot geographical dis- function performance statistics generated from the cal ibration set tances. Plots that are close to each other tend to have similar itself by leave-one-out cross-validation may have been over- species frequencies and climatic conditions. optimistic (Telford and Birks, 2005, 2009; Telford, 2006). To avoid this problem, we would have had to base the calculation of 4. Discussion performance statistics on a spatially independent test set, not on the calibration set. While construction of such a test set would not We created a calibration set consrstmg of contemporary have been feasible, WA-PLS models are fairly robust to spatial vascular plant species abundances (both quadrat and subplot autocorrelation. frequencies) and climate data for 107 FlA plots in Oregon and Washington. Through WA-PlS, we created six transfer functions to 4.1. Accounting for transfer function performance enable the prediction of MJAT(mean January temperature), MJUT (mean July temperature), and MANPt (mean annual precipitation Both data quality and statistical considerations likely are transformed to natural logarithms) from quadrat or subplot responsible for the good predictive abilities of our transfer frequencies measured, in the future, on those plots. We assessed functions. the predictive abilities of the transfer functions (all are two- We computed the vascular plant species frequencies from component WA-PlS models) by leave-one-out cross-validation. inventory data for FIAPhase 3 plots. An elaborate quality assurance We expected the transfer functions based on subplot frequen- program helps to maintain the veracity of plot data. This includes cies to prove superior, because the subplot frequencies database the training, testing, and certification of experienced botanists as contains more species (698) than the quadrat frequencies database well as field audits (Pollard, nd: Schulz, nd). Gray and Azuma (538). However, the choice of species database had little effect on (2005) reported the outcome of an assessment of data collector performance. precision involving two botanists (working independently) and 48 The coefficients of determination (1'2) linked to the transfer FIA Phase 3 plots from Oregon sampled in 2000: The botanists functions for MJAT and MANPt are very high (0.85 or greater). agreed on species identifications 71% of the time (Sorensen's index Recent palaeoecological studies with comparable statistical of similarity) within subplots and 67%of the time within quadrats. methods (Lu et al., 2006; Shen et al., 2006; Finsinger et al., The researchers concluded that differences of opinion generally 2007) found similar values for mean January temperature and concerned congeneric species or the missing of rare plants, and (untransformed) mean annual precipitation transfer functions. Our that repeatability was comparable to results of other botanical MJUT transfer functions are much less powerful (r2 = 0.63-0.64). A studies. The species data we used here sampled 20-30% of the plot 2 large range of r statistics (0.33-0.84) characterized mean July locations potentially examined for these states over a full temperature transfer functions in recent palaeoecological inves- measurement cycle. A larger species dataset possibly would tigations with analogous methods (Rosen et al., 2003; Shen et al., contain fewer zeros and reveal stronger relationships with 2006; Finsinger et al., 2007). RMSEPs expressed as percentages of estimated climate. variable ranges, are similar to, or slightly better than, recently For each plot, we obtained climate data from the Daymet model published values for our temperature transfer functions (8.31- database. Thornton et al. (1997) carried out a leave-one-out cross- 10.12%). Absolute errors (back-transformed RMSEPs that depend validation analysis to evaluate the predictive ability of Daymet. on the position along the gradient) expressed as percentages of the They compared modeled daily maximum temperature, daily range of mean annual precipitation lie between about 2%and 33% minimum temperature, and daily precipitation with actual for our MANPt transfer functions. The corresponding relative measurements over 1 year (1989) at nearly 500 sites in the inland errors (back-transformed RMSEPs that are independent of gradient northwestern USA (including southeastern Washington and position) are close to 30%. Due to transformation, the RMSEPs northeastern Oregon). Mean absolute error of prediction, MAE, associated with our MANPt transfer functions are not readily and bias were 0.7 and -0.1 o C, respectively, for the annual mean of comparable to published values for mean annual precipitation, daily maximum temperature. MAE and bias were 1.2 and 0.1oC although our RMSEPs for the dry end of the precipitation gradient respectively, for the annual mean of daily minimum temperature. are appreciably lower than recently published percentages for MAE and bias were 134 and -21 mm, respectively, for annual untransformed mean annual precipitation transfer functions precipitation. These variables and performance statistics differ (7.25-8.00%; Lu et al., 2006; Shen et al., 2006). We had expected from those used in our study. However, according to the authors, our transfer functions to significantly outperform the paleoecolog- the prediction errors associated with the temperature variables are ical ones, as our frequency data represent direct measures of plot- comparable to published values for other climate models. Daymet successfully predicted the occurrence of daily precipitation 83.3% of the time, a seemingly commendable rate. The transformation of mean annual precipitation to natural logarithms is responsible, probably primarily, for the good performances of our transfer functions for MANPt (very high 1'2 values; very low RMSEPs at the dry end of the gradient). We obtained transfer functions with remarkably poor performance statistics after applying WA-PLS to un transformed mean annual precipitation data (MANP). The unsatisfactory outcomes included the one-component models, which suggest that the advantages of transformation emerge when determining weighted averages for the original environmental variable. Prior to transformation, MANP was badly skewed, so we expected many species response curves to exhibit considerable skewness as well. In fact, when we fitted HOF models to quadrat frequencies data and MANP (we used the same methods as described above), the symmetrical unimodal curve (Model IV) still was the dominant type, but the number of species responses characterized by it declined substantially. In contrast, the monotonic increasing or decreasing curve (Model II) became much more common. When response curves are not symmetrical, weighted averaging will provide poor estimates of species optima, which leads to a defective WA-PLS transfer function. Transformation of MANP reduced the distortion by reducing or eliminating the skewness in many species response curves. The logic here is somewhat akin to that used to explain the effects of sampling evenness on weighted average estimates of species optima (Ter Braak and looman, 1986). Logarithmic or square-root transformation of environmental data is no guarantee of good results. Such transformations of our temperature data failed to yield improved transfer functions (the transformation of temperature data makes little ecological sense anyway). Clearly, simulation studies are warranted to identify the consequences of environmental data transformation on transfer function construc- tion by WA-PLS.

4.2. Capability of detecting future climate change

The primary objective of this research is to develop a set of transfer functions that can provide high quality quantitative information about future local climatic conditions based on future measurements of vascular plant species frequencies on FIAPhase 3 plots. The usefulness of our transfer functions rests upon thei r abilities to detect short-term changes in climate. The magnitude of its predictive error (i.e., RMSEPand maximum bias) and the pace of climate change expected over the coming decades establish the apparent resolution of each transfer function. The actual predictive power of a transfer function also hinges on the rates of response of species frequencies to climate change (i.e.. the relative lengths of lags). As reported by Mote et al. (2008), the Climate Impacts Group at the University of Washington evaluated 20 global simulation models to develop predictions of future climate (30-year means for selected variables) in the Pacific Northwest, USA, which includes Oregon and Washington. Each of the models had produced simulations for scenarios of low (scenario B1) and high (scenario AlB) projected changes in future concentrations of greenhouse gases, which promote atmospheric warming, and sulfate aerosols, which partly ameliorate warming. For a given emissions scenario and climate variable, the Climate Impacts Group determined the amount of change, specific to the Pacific Northwest, predicted by each of the 20 models relative to an average for 1970-1999. They also computed a weighted average of the 20 predictions (models that performed well in simulating twentieth century climate received greater weight). Fig. 4 facilitates the assessment of the apparent abilities of our transfer functions to detect the climate changes expected for Oregon and Washington by global models for the period 2010- Our transfer functions possess good predictive abilities 2039. Each graph shows the observed and predicted (by two- compared to those reported in the literature. Both data quality component WA-PLS model for quadrat frequencies) climate values and statistical considerations (viz., the logarithmic transformation for plots from the calibration set (1980-1997) as well as the lowest, of mean annual precipitation) probably are responsible. The highest, and weighted average model predictions for 2010-2039. discovery of spatial autocorrelation in the calibration set may We obtained the data generated by the Climate Impacts Group portend some inflation of the apparent predictive abilities of our at http://cses.washington.edu/cig/fpt/08scensumdata.shtml. This transfer functions. source does not offer temperature predictions by month. Our MJUTtransfer functions may become useful soon in view of Therefore, we based predicted changes in MJATon those for mean the rapid rate of increase in summer temperatures expected during winter (December, January, February) temperature, and predicted the first few decades of this century by some global simulation changes in MJUT on those for mean summer (June, July, August) models. Our MJATtransfer functions may be able to track changes temperature. Since scenarios B1 and A1B give similar results prior in mean January temperature by the 2040s. Our MANPt transfer to the 2050s, we averaged the weighted averages calculated under functions may begin detecting changes in mean annual precipita- these opposing propositions. To obtain future climate values, we tion by the 2070s. Our MJATand MANPt transfer functions can be simply added the predicted changes to the estimated Daymet used soon to verify relative climatic stability, iflags are very short. values for all plots from the calibration set (1980-1997). Note that Our transfer functions, whether based on quadrat frequencies we assumed that all plots experience the same amount of change or on subplot frequencies, successfully predict the overall climatic over a given interval of time. This seems reasonable, as twentieth trends in the calibration set. However, because the predicted and century patterns of change in temperature and precipitation were observed values sometimes deviate substantially for individual roughly uniform across the Pacific Northwest (Mote, 2003). plots, our transfer functions are appropriate for monitoring Our transfer functions will possess sufficient resolution for regional climatic patterns, not for assessing climate change at detecting the climatic trends expected to characterize Oregon and specific plots. Washington during the first part of this century only if the amount of change taking place significantly exceeds the sizes of the Aclmowledgements predictive errors estimated by leave-one- out cross-validation, and if the lags of at least a fair number of species are no more than J. Morefield, J. Pijoan, C. Grenz, and S. Butler collected field data. several decades in length. As is evident from Fig. 4, the errors S. Sundberg and K. Mitchell (Oregon State University Herbarium) associated with the MJATtransfer functions are great enough that identified voucher specimens. R. S. (reagan edited the References even the highest amount of change predicted for the period 2010- section. J. S. Fried and two anonymous reviewers commented on 2039 would be indistinguishable from current climatic conditions. earlier versions of the manuscript. On the other hand, the MJUTtransfer functions have the abilities to detect change under the highest projection for 2010-2039. The References results of 20 global models indicate that only slight changes in mean annual precipitation will occur during the early part of the Andrewartha, H.G.. Birch. LC, 1954. The Distribution and Abundance of Animals. twenty-first century. If the predictions are correct, our MANPt University of Chicago Press, Chicago. Belsley, D.A.. Kuh, E.. Welsch, RE., 1980. Regression Diagnostics: Identifying Influ- transfer functions will have no new climatic trend to monitor. ential Data and Sources of Collinearity, Wiley, New York. Our MJUT transfer functions may become useful soon in view of Birks, H.J.B.. 1995. Quantitative palaeoenvironmental reconstructions. In: Maddy, the rapid rate of increase in summer temperatures expected during D., Brew, J.S. (Eds.), Statistical Modelling of Quaternary Science Data, Technical the first few decades of this century by some global simulation Guide, vol. 5. Quaternary Research Association, Cambridge, pp. 161-254. Birks, H.J.B., 1998. Numerical tools in palaeolimnology-progress, potentialities, and models. The higher model predictions indicate that our MJAT problems. j. Palaeolimnol. 20, 307-332. transfer functions may be able to track changes in mean January Cleland, D.T.. Freeouf, JA. Keys Jr .. J.E.. Nowacki, G.J.. Carpenter, C.A., McNab, W.H .. temperature by the 2040s. In addition, the higher model 2007. Ecological Subregions: Sections and Subsections for the Conterminous United States [I :3,500,000] [CD-ROM]. USDA Forest Service, Washington, D.c.. predictions suggest that our MANPt transfer functions may begin Davis, M.B., Botkin, D.B.. 1985. Sensitivity of the fossil pollen record to sudden detecting changes in mean annual precipitation by the 2070s. Of climatic change. Quat. Res. 23, 327-340. course, our MJAT and MANPt transfer functions may prove useful Enright, J.T., 1976. Climate and population regulation: the biogeographers dilemma. Oecologia 24, 295-310. at the present time to verify relative climatic stability (assuming Finsinger. W., Heiri, 0., Valsecchi, V., Tinner, W .. Lotter. A.F.. 2007. Modern pollen that lags are very short). assemblages as climate indicators in southern Europe. Global Ecol. Biogeogr. 16, Fig. 4 attests to the general correlation between observed 567-582. Franklin, J.F .. Dyrness, C.T., 1988. Natural Vegetation of Oregon and Washington. climate values and those predicted by our transfer functions. Oregon State University Press, Corvallis. Because the predicted values sometimes deviate substantially Garbolino, E., De Ruffray, P., Brisse. H.. Grandjouan, G.. 2007. Relations entre planres from the observed values for individual plots, our transfer et climats en France: etalonnage de 1874 bio-indicareurs. C. R Biologies 330, 159-170. functions are appropriate for monitoring climatic trends over Gray, A.N., Azuma, D.L, 2005. Repeatability and implementation of a forest vegeta- the entire Pacific Northwest or large regions within it, not for tion indicator. Ecol. Indicators 5, 57-71. assessing climate change at individual plots. Hartmann, H.C., Pagano, T.C., Sorooshian, S., Bales, R, 2002. Confidence builders: evaluating seasonal climate forecasts from user perspectives. Bull. Am. Meteorol. Soc. 83, 683-698. 5. Conclusions Hill, M.O .. Gauch Jr., H.G., 1980. Detrended correspondence analysis: an improved ordination technique. Vegetatio 42, 47-58. Howe, S., Webb III, T.. 1983. Calibrating pollen data in climatic terms: improving the To monitor future climate changes in Oregon and Washington, methods. Quat. Sci. Rev. 2, 17-51. we exploited the link between vascular plant species abundance Huisman, J .. Olff, H., Fresco, L.F.M.. 1993. A hierarchical set of models for species and climate: We assembled a calibration set consisting of present- response analysis. J. Veg. Sci. 4, 37-46. day quadrat and subplot species frequencies and climate estimates IPCC, 2007. Summary for policymakers. In: Solomon, S.. Qin. D., Manning, M., Chen, Z., Marquis, M .. Averyt, K.B., Tignor, M.. Miller, H.1.. (Eds.), Climate Change for 107 FIAPhase 3 plots. We used WA-PLS to calculate six transfer 2007: The Physical Science Basis. Contribution of Working Group I to the Fourth functions that will permit the prediction of MJAT (mean January Assessment Report of the Intergovernmental Panel on Climate Change. Cam- temperature), MJUT (mean July temperature), and MANPt (mean bridge University Press, Cambridge. Juggins. S., 2007. C2 (version 1.5.0), a Microsoft Windows Program for the Analysis annual precipitation transformed to natural logarithms) from and Visualization of Ecological and Palaeoenvironmental Data. Available at future quadrat or subplot frequency measurements. http://www.campus.ncl.ac.u k/staff/Stephen.Juggins. Karlsen, S.R., Elvebakk, A.. 2003. A method using indicator plants to map local Schulz, B.K.. nd. Forest Inventory and Analysis: Vegetation Indicator. FIA Fact Sheet climatic variation in the Kangerlussuaq/Scoresby Sund area, East Greenland. J. Series. Available at http://fia.fs.fed.us/library/fact-sheets/· Biogeogr. 30, 1469- 1491. Schulz, B.K.. Bechtold, W.A .. Zarnoch, S.J., 2009. Sampling and Estimation Proce- Krebs, Cj., 1985. Ecology: The Experimental Analysis of Distribution and dures for the Vegetation Diversity and Structure Indicator. PNW-GTR-781. Abundance, 3rd ed. Harper and Row, New York. USDA Forest Service, Pacific Northwest Research Station, Portland, OR. Leps, J., Smilauer, p.. 2003. Multivariate Analysis of Ecological Data Using CANOCO. Shen. c.. Liu, K... Tang, L., Overpeck, J.T .. 2006. Quantitative relationships between Cambridge University Press, Cambridge, UK. modern pollen rain and climate in the Tibetan Plateau. Rev. Palaeobot. Palynol. Li, Y., Xu, Q. Liu. J .. Yang, X.. Nakagawa, T., 2007. A transfer-function model 140,61-77. developed from an extensive surface-pollen data set in northern China and Smith, A.G.. 1965. Problems of inertia and threshold related to postglacial habitat its potential for palaeoclimate reconstructions. Holocene 17, 897-905. changes. Proc, R. Soc. Lond . Ser. B 61, 331-342. Lischke, H.. Lotter, A.F.. Fischlin, A., 2002. Untangling a Holocene pollen record with Smith, S.D.. Bunting, S.E., Hironaka, M. 1986. Sensitivity of frequency plots for forest model simulations and independent climate data. Ecol. Model. 150, 1-21. detecting vegetation change. Northwest Sci. 60, 279-286. Lu, H., Wu, N.. Yang, X., Jiang, H.. Liu, K.. Liu, T., 2006. Phytoliths as quantitative Stapanian, M.A.. Cline, S.P., Cassell, D.L., 1997. Evaluation of a measurement method indicators for the reconstruction of past environmental conditions in China I: for forest vegetation in a large-scale ecological survey. Environ. Monit. Assess, phytolith-based transfer functions, Quat. Sci. Rev. 25, 945-959. 45, 237-257. Mantei. N.. 1967, The detection of disease clustering and generalized regression St. Jacques, J., Cumming, B,F.. Smol, J.P .. 2008. A pre-European settlement pollen- approach, Cancer Res. 27, 209-220. climate calibration set for Minnesota, USA: developing tools for palaeoclimatic Marquardt, D.W .. 1970. Generalized inverses, ridge regression, biased linear esti- reconstructions. J. Biogeogr. 35, 306-324. mation, and nonlinear estimation. Technometrics 12 (591). 605-607. Svenning, J-C; Normand, S., Skov, F.. 2008. Postglacial dispersal limitation of McCune, B.. Grace, J.B .. 2002. Analysis of Ecological Communities. MjM Software, widespread forest plant species in nemoral Europe. Ecography 31, 316-326. Gleneden Beach, Oregon, USA. Telford, R.J.. 2006. Limitations of dinoflagellate cyst transfer functions. Quat. Sci. McCune, B.. Mefford, M.J., 2005. PC-ORD (version 5.0), a Microsoft Windows Rev. 25, 1375-1382. Program for the Multivariate Analysis of Ecological Data. MjM Software, Gle- Telford, R.J.. Birks, H.J.B., 2005. The secret assumption of transfer functions: pro- neden Beach, Oregon, USA. blems with spatial autocorrelation in evaluating model performance. Quat. Sci. Miles, E.I.., Snover, A.K.. Whitely Binder, L.C., Sarachik, E.S., Mote, P.W .. Mantua, N.. Rev. 24, 2173-2179. 2006. An approach to designing a national climate service. Proc. Natl. Acad. Sci. Telford, R.J, Birks, H,j.B., 2009. Evaluation of transfer functions in spatially struc- 103, 19616-19623. tured environments. Quat. Sci, Rev. 28, 1309-1316. Miller, P.A.. Giesecke, T., Hickler, T.. Bradshaw, R.H.W .. Smith, B., Seppa, H. Valdes, Ter Braak, C.J.F; 1986. Canonical correspondence analysis: a new eigenvector P.J.. Sykes, M.T.. 2008. Exploring climatic and biotic controls on Holocene technique for multivariate direct gradient analysis. Ecology 67, 1167-1179. vegetation change in Fennoscandia. J. Ecol. 96, 247-259. Ter Braak, Cj.F; 1987. The analysis of vegetation-environment relationships by Mote, P.W .. 2003. Trends in temperature and precipitation in the Pacific Northwest canonical correspondence analysis. Vegetatio 69, 69-77. during the twentieth century. Northwest Sci. 77, 271-282. Ter Braak, E.J.F. 1995. Non-linear methods for multivariate statistical calibration Mote, P.W., Salathe. E.P., Duliere, V., jump, E.. 2008. Scenarios of Future Climate and their use in palaeoecology: a comparison of inverse (k-nearest neighbours), Change for the Pacific Northwest. Report prepared by the Climate Impacts partial least squares and weighted averaging partial least squares and classical Group, Center for Science in the Earth System Joint Institute for the Study of the approaches. Chemomet. Intell. Lab, Syst. 28, 165-180. Atmosphere and Oceans, University of Washington, Seattle. Available at http:// Ter Braak, C.J.F .. Juggins, S.. 1993. Weighted averaging partial least squares regres- cses.washington.edu/db/pubs/abstract578.shtml. sion (WA-PLS): an improved method for reconstructing environmental vari- Neumann, F.. Schoelzel, C; Litt, T.. Hense, A.. Stein, M.. 2007, Holocene vegetation ables from species assemblages. Hydrobiologia 269/270, 485-502, and climate history of the northern Golan heights (Near East). Veg. Hist. Ter Braak, E.J.F .. Looman, E.W,N., 1986, Weighted averaging, logistic regression and Archaeobot. 16, 329-346. the Gaussian response model. Vegetatio 65, 3-11. Okland, R.H.. 1986. Reseating of ecological gradients. II. The effect of scale on Ter Braak, C.J.F .. Prentice, I.C, 1988. A theory of gradient analysis. Adv. Ecol. Res. 18, symmetry of species response curves. Nordic J. Bot. 6, 661-670. 271-317. Oksanen. J., Minchin, P.R.. 2002. Continuum theory revisited: what shape are species Thornton, P.E.. Running, S.W .. White, M.A.. 1997. Generating surfaces of daily responses along ecological gradients? Ecol. Model. 157, 119-129. meteorological variables over large regions of complex terrain. J. Hydrol. Palmer, M.W .. 1993. Putting things in even better order: the advantages of canonical 190,214-251. correspondence analysis. Ecology 74, 2215-2230. Valsecchi, V., Finsinger, W., Tinner, W., Ammann, B.. 2008. Testing the influence of Payette, S., 2007. Contrasted dynamics of Northern Labrador tree lines caused by climate, human impact and fire on the Holocene population expansion of Fagus climate change and migrational lag. Ecology 88, 770-780. sylvatica in the southern Prealps (Italy). Holocene 18,603-614. Pollard, J. nd. Forest Inventory and Analysis: Quality Assurance. FIA Fact Sheet Webb III, T.. 1986, Is vegetation in equilibrium with climate? How to interpret late- Series. Available at http://fia.fs.fed.us/library/fact-sheets/. Quaternary pollen data. Vegetatio 67, 75-91. Prentice, I.C, Monserud, R.A.. Smith, T.M., Emanuel, W.R .. 1993. Modeling large- Wilmshurst, J.M., McGlone, M.S., Leathwick, J.R., Newnharn, R,M .. 2007. A pre- scale vegetation dynamics, In: Solomon, A.M., Shugart, H.H. (Eds.), Vegetation deforestation pollen-climate calibration model for New Zealand and quantita- Dynamics and Global Change. Chapman and Hall, New York, pp. 235-250. tive temperature reconstructions for the past 18,000 years BP.J. Quat. Sci. 22, Rosen, P.. Segerstrom, U.. Eriksson, L.. Renberg, I.. 2003. Do diatom, chironomid, and 535-547. pollen records consistently infer Holocene July air temperature? A comparison Woodward, F.I., 1987. Climate and Plant Distribution, Cambridge University Press, using sediment cores from four alpine lakes in northern Sweden. Arctic Cambridge. Antarctic Alpine Res. 35, 279-290. Woodward, F.I., 1992. Predicting plant responses to global environmental change. Sachs, H.M .. Webb III, T., Clark, D.R., 1977. Paleoecological transfer functions. Ann. New Phytol. 122,239-251, Rev. Earth Planet. Sci. 5, 159-178. Woodward, F.I., Williams, B.G.. 1987. Climate and plant distribution at global and SAS Institute Inc., 2000. JMP (version 4.0.2). Cary, Ne. 1989-2007. local scales, Vegetatio 69,189-197.